Re: [PATCH] 2.6.6 invalid usage of GFP_DMA in drivers/scsi/pluto.c

[Yury Umanets <torque@ukrpost.net>]

On Tue, 2004-06-08 at 13:17, Andrew Morton wrote:
> Yury Umanets <torque@ukrpost.net> wrote:
> >
> > Hello Andrew, guys,
> > 
> > Found this, what seems to be an invalid usage of GFP_DMA flag. Is this
> > patch okay?
> 
> Nope.
> 
> GFP_DMA means "from the lower 16MB of memory".  It's needed for crufty old
> eisa hardware which only does 24-bit DMA.
This is clear. And sure, it shows that caller wants some memory from DMA
zone if it is possible.

What I mean is that, GFP_DMA seems to be not full specifier and only
flag which can be used along with something else. 

Say gfp mask roughly consist of two kinds on instructions:
* what memory zone to use - zone specifier.
* how to behave during allocation (IO, sleep, etc) - behavior specifier.

My concern is will it behave correctly with only GFP_DMA? It seems to
behave like do not use emergency pools and do not wait at the same time
what is risky. Is that right?

>   It's meaningless to OR this with
> GFP_KERNEL.

> 
> However it's a bit odd that GFP_DMA implies !__GFP_WAIT.  It would be valid
> to hunt down GFP_DMA users who should really be using GFP_DMA|__GFP_WAIT,

Seems that all GFP_DMA users use it with __GFP_WAIT or GFP_KERNEL. And
I'd prefer to add __GFP_WAIT here also.


> but this stuff is so old and crufty I'd be inclined to leave it all alone.
> 
> 
> > 
> > --- ./linux-2.6.6/drivers/scsi/pluto.c	Mon May 10 05:32:27 2004
> > +++ ./linux-2.6.6-modified/drivers/scsi/pluto.c	Tue Jun  8 11:26:07 2004
> > @@ -117,7 +117,8 @@ int __init pluto_detect(Scsi_Host_Templa
> >  #endif
> >  			return 0;
> >  	}
> > -	fcs = (struct ctrl_inquiry *) kmalloc (sizeof (struct ctrl_inquiry) *
> > fcscount, GFP_DMA);
> > +	fcs = (struct ctrl_inquiry *) kmalloc (sizeof (struct ctrl_inquiry) *
> > fcscount, 
> > +			GFP_KERNEL | GFP_DMA);
> >  	if (!fcs) {
> >  		printk ("PLUTO: Not enough memory to probe\n");
> >  		return 0;
> > 
> 

> Your patch is wordwrapped and uses weird headers (please omit the leading
> ./ from the pathnames).
Sorry, next will use another mailer.

-- 
umka